Physarum Powered Differentiable Linear Programming Layers and Applications

Consider a learning algorithm, which involves an internal call to optimization routine such as generalized eigenvalue problem, cone programming problem or even sorting. Integrating method layers within trainable deep network in numerically stable way is not simple – for instance, only recently, strategies have emerged eigendecomposition and differentiable We propose efficient solver general linear problems can be used plug play manner neural networks layer. Our development inspired by fascinating but widely link between dynamics of slime mold (physarum) mathematical schemes steepest descent. describe our demonstrate the use video object segmentation task meta-learning few-shot learning. review relevant known results provide technical analysis describing its applicability cases. performs comparably with customized projected gradient descent on first outperforms very recently proposed CVXPY second task. Experiments show that converges quickly without need feasible initial point. Interestingly, scheme easy implement easily serve whenever procedure needs fast approximate solution LP, larger network.